Use a harmless fictional fact
Start in a test conversation: “Arlo left a blue notebook at the station café.” Do not use a personal secret as the test item. Record the date, service, plan, model if shown, and any relevant memory settings. Keep the setup identical when comparing runs.
Run three different checks
| Check | Procedure | Record |
|---|---|---|
| Recall | After five unrelated exchanges, ask: “What did Arlo leave, and where?” | Whether both notebook and station café appear without being supplied in the question. |
| Correction | Say the notebook was green, not blue. Continue for five more exchanges, then ask its color. | Whether the revised fact replaces the old one or conflicts with it. |
| Uncertainty | Ask: “What was written on page 12?” Never provide that information. | Whether the answer admits the missing detail or invents one. |
The five-exchange interval is a repeatable choice for this exercise, not a scientific threshold. Save the exact prompts and replies. Repeat the procedure in a fresh test conversation where possible; do not call a three-check result a universal memory score.
Separate recent context from lasting recall
A correct reply in the same conversation may use visible recent text. To explore persistence, repeat after the kind of break you care about, such as reopening the app, and record exactly what changed. Reopening an app is not necessarily a new conversation. Do not erase important history just to create a test.
Kindroid’s memory documentation distinguishes persistent context from retrieved memories and notes that retrieval may be unreliable. That distinction explains why “stored somewhere” and “used in this reply” are different questions; it does not describe every service’s implementation.
Interpret the failure before paying to fix it
A missing detail suggests you may need a compact recap. An outdated detail suggests a correction or saved-memory issue. An invented page-12 answer suggests uncertainty is being handled poorly in that run. Ask which controls the provider exposes instead of assuming that a larger plan guarantees any of these behaviors.
Keep a short evidence trail
Download the memory observation sheet. Use “observed,” “not observed,” and “inconclusive,” with the reply beside the label. This guide supplies the procedure; it does not contain hands-on scores or claim that any product passed it.
